Spaces:
Build error
Build error
Update app.py
Browse files
app.py
CHANGED
|
@@ -2,27 +2,19 @@ import os
|
|
| 2 |
import sys
|
| 3 |
import subprocess
|
| 4 |
|
| 5 |
-
# --- УСТАНОВКА
|
| 6 |
-
# Устанавливаем версию с поддержкой Vision (CPU)
|
| 7 |
try:
|
| 8 |
-
from llama_cpp import Llama
|
| 9 |
-
|
| 10 |
-
print("Библиотека llama-cpp-python и Qwen2VLChatHandler найдены.")
|
| 11 |
except ImportError:
|
| 12 |
-
print("Установка
|
|
|
|
| 13 |
subprocess.check_call([
|
| 14 |
sys.executable, "-m", "pip", "install",
|
| 15 |
-
"llama-cpp-python>=0.3.
|
| 16 |
"--extra-index-url", "https://abetlen.github.io/llama-cpp-python/whl/cpu"
|
| 17 |
])
|
| 18 |
-
|
| 19 |
-
from llama_cpp import Llama
|
| 20 |
-
# Пытаемся импортировать хендлер после установки
|
| 21 |
-
try:
|
| 22 |
-
from llama_cpp.llama_chat_format import Qwen2VLChatHandler
|
| 23 |
-
except ImportError:
|
| 24 |
-
print("ВАЖНО: Qwen2VLChatHandler не найден. Возможно, версия библиотеки старая.")
|
| 25 |
-
Qwen2VLChatHandler = None
|
| 26 |
|
| 27 |
import gradio as gr
|
| 28 |
from huggingface_hub import hf_hub_download
|
|
@@ -31,45 +23,96 @@ import io
|
|
| 31 |
import re
|
| 32 |
from PIL import Image
|
| 33 |
|
| 34 |
-
#
|
| 35 |
REPO_ID = "mradermacher/VisualQuality-R1-7B-GGUF"
|
| 36 |
MODEL_FILENAME = "VisualQuality-R1-7B.Q8_0.gguf"
|
| 37 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
llm = None
|
| 39 |
|
| 40 |
def load_model():
|
| 41 |
global llm
|
| 42 |
if llm is None:
|
| 43 |
print(f"Загрузка модели {MODEL_FILENAME}...")
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
return llm
|
| 65 |
|
| 66 |
def process_image(image):
|
| 67 |
-
# Ресайз
|
| 68 |
-
|
| 69 |
-
if max(image.size) >
|
| 70 |
-
|
| 71 |
-
new_size = (int(image.size[0] * ratio), int(image.size[1] * ratio))
|
| 72 |
-
image = image.resize(new_size, Image.Resampling.LANCZOS)
|
| 73 |
|
| 74 |
buffered = io.BytesIO()
|
| 75 |
image = image.convert("RGB")
|
|
@@ -81,15 +124,15 @@ def evaluate_image(image, progress=gr.Progress()):
|
|
| 81 |
return "Пожалуйста, загрузите изображение.", ""
|
| 82 |
|
| 83 |
try:
|
| 84 |
-
progress(0.1, desc="
|
| 85 |
model = load_model()
|
| 86 |
|
| 87 |
-
progress(0.
|
| 88 |
-
|
| 89 |
-
|
| 90 |
|
| 91 |
system_prompt = "You are doing the image quality assessment task."
|
| 92 |
-
|
| 93 |
"What is your overall rating on the quality of this picture? "
|
| 94 |
"The rating should be a float between 1 and 5, rounded to two decimal places, "
|
| 95 |
"with 1 representing very poor quality and 5 representing excellent quality. "
|
|
@@ -101,15 +144,16 @@ def evaluate_image(image, progress=gr.Progress()):
|
|
| 101 |
{
|
| 102 |
"role": "user",
|
| 103 |
"content": [
|
| 104 |
-
{"type": "image_url", "image_url": {"url":
|
| 105 |
-
{"type": "text", "text":
|
| 106 |
]
|
| 107 |
}
|
| 108 |
]
|
| 109 |
|
| 110 |
full_response = ""
|
| 111 |
-
print("
|
| 112 |
|
|
|
|
| 113 |
stream = model.create_chat_completion(
|
| 114 |
messages=messages,
|
| 115 |
max_tokens=1024,
|
|
@@ -123,22 +167,23 @@ def evaluate_image(image, progress=gr.Progress()):
|
|
| 123 |
if "content" in delta and delta["content"]:
|
| 124 |
content = delta["content"]
|
| 125 |
full_response += content
|
| 126 |
-
yield full_response, "
|
| 127 |
|
| 128 |
-
#
|
| 129 |
score_match = re.search(r'<answer>\s*([\d\.]+)\s*</answer>', full_response)
|
| 130 |
-
final_score = score_match.group(1) if score_match else "
|
| 131 |
|
| 132 |
yield full_response, final_score
|
| 133 |
|
| 134 |
except Exception as e:
|
| 135 |
-
|
| 136 |
-
print(
|
| 137 |
-
yield
|
| 138 |
|
| 139 |
-
|
|
|
|
| 140 |
gr.Markdown("# 👁️ VisualQuality-R1 (Qwen2-VL)")
|
| 141 |
-
gr.Markdown("Оценка качества
|
| 142 |
|
| 143 |
with gr.Row():
|
| 144 |
with gr.Column():
|
|
@@ -147,7 +192,7 @@ with gr.Blocks(title="VisualQuality-R1") as demo:
|
|
| 147 |
|
| 148 |
with gr.Column():
|
| 149 |
output_score = gr.Label(label="Оценка")
|
| 150 |
-
output_text = gr.Textbox(label="CoT (
|
| 151 |
|
| 152 |
run_btn.click(evaluate_image, inputs=[input_img], outputs=[output_text, output_score])
|
| 153 |
|
|
|
|
| 2 |
import sys
|
| 3 |
import subprocess
|
| 4 |
|
| 5 |
+
# --- ПРОВЕРКА И УСТАНОВКА БИБЛИОТЕКИ ---
|
|
|
|
| 6 |
try:
|
| 7 |
+
from llama_cpp import Llama, LlamaChatCompletionHandler
|
| 8 |
+
print("Библиотека llama-cpp-python найдена.")
|
|
|
|
| 9 |
except ImportError:
|
| 10 |
+
print("Установка llama-cpp-python (CPU)...")
|
| 11 |
+
# Принудительно ставим 0.3.16 или новее с поддержкой CPU
|
| 12 |
subprocess.check_call([
|
| 13 |
sys.executable, "-m", "pip", "install",
|
| 14 |
+
"llama-cpp-python>=0.3.16",
|
| 15 |
"--extra-index-url", "https://abetlen.github.io/llama-cpp-python/whl/cpu"
|
| 16 |
])
|
| 17 |
+
from llama_cpp import Llama, LlamaChatCompletionHandler
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
|
| 19 |
import gradio as gr
|
| 20 |
from huggingface_hub import hf_hub_download
|
|
|
|
| 23 |
import re
|
| 24 |
from PIL import Image
|
| 25 |
|
| 26 |
+
# Конфигурация
|
| 27 |
REPO_ID = "mradermacher/VisualQuality-R1-7B-GGUF"
|
| 28 |
MODEL_FILENAME = "VisualQuality-R1-7B.Q8_0.gguf"
|
| 29 |
|
| 30 |
+
# === ГЛАВНЫЙ ФИКС: СВОЙ ОБРАБОТЧИК ДЛЯ QWEN2-VL ===
|
| 31 |
+
# Мы не зависим от встроенных классов, а пишем свой.
|
| 32 |
+
class CustomQwen2VLHandler(LlamaChatCompletionHandler):
|
| 33 |
+
def __init__(self, clip_model_path=None, verbose=False):
|
| 34 |
+
self.clip_model_path = clip_model_path
|
| 35 |
+
self.verbose = verbose
|
| 36 |
+
|
| 37 |
+
def __call__(self, llama: Llama, messages, functions=None, function_call=None, tools=None, tool_choice=None, **kwargs):
|
| 38 |
+
# 1. Формируем промпт вручную с правильными тегами
|
| 39 |
+
prompt = ""
|
| 40 |
+
images = []
|
| 41 |
+
|
| 42 |
+
for message in messages:
|
| 43 |
+
role = message["role"]
|
| 44 |
+
content = message["content"]
|
| 45 |
+
|
| 46 |
+
# Начало сообщения
|
| 47 |
+
prompt += f"<|im_start|>{role}\n"
|
| 48 |
+
|
| 49 |
+
if isinstance(content, str):
|
| 50 |
+
prompt += content
|
| 51 |
+
elif isinstance(content, list):
|
| 52 |
+
for part in content:
|
| 53 |
+
if part["type"] == "text":
|
| 54 |
+
prompt += part["text"]
|
| 55 |
+
elif part["type"] == "image_url":
|
| 56 |
+
# Теги для Qwen2-VL: Vision Start -> Pad -> Vision End
|
| 57 |
+
prompt += "<|vision_start|><|image_pad|><|vision_end|>"
|
| 58 |
+
|
| 59 |
+
# Извлекаем байты из base64 для передачи в C++ слой
|
| 60 |
+
try:
|
| 61 |
+
image_url = part["image_url"]["url"]
|
| 62 |
+
if "base64," in image_url:
|
| 63 |
+
base64_data = image_url.split("base64,")[1]
|
| 64 |
+
image_bytes = base64.b64decode(base64_data)
|
| 65 |
+
images.append(image_bytes)
|
| 66 |
+
except Exception as e:
|
| 67 |
+
print(f"Ошибка декодирования картинки: {e}")
|
| 68 |
+
|
| 69 |
+
# Конец сообщения
|
| 70 |
+
prompt += "<|im_end|>\n"
|
| 71 |
+
|
| 72 |
+
# Добавляем триггер для ответа ассистента
|
| 73 |
+
prompt += "<|im_start|>assistant\n"
|
| 74 |
+
|
| 75 |
+
if self.verbose:
|
| 76 |
+
print(f"=== SENDED PROMPT ({len(prompt)} chars) ===")
|
| 77 |
+
print(prompt[:200] + "..." if len(prompt) > 200 else prompt)
|
| 78 |
+
print(f"=== IMAGES: {len(images)} ===")
|
| 79 |
+
|
| 80 |
+
# Возвращаем кортеж (prompt, images), который понимает llama.cpp
|
| 81 |
+
return prompt, images
|
| 82 |
+
|
| 83 |
llm = None
|
| 84 |
|
| 85 |
def load_model():
|
| 86 |
global llm
|
| 87 |
if llm is None:
|
| 88 |
print(f"Загрузка модели {MODEL_FILENAME}...")
|
| 89 |
+
try:
|
| 90 |
+
model_path = hf_hub_download(repo_id=REPO_ID, filename=MODEL_FILENAME)
|
| 91 |
+
|
| 92 |
+
# Инициализируем НАШ кастомный хендлер
|
| 93 |
+
# clip_model_path указываем на тот же файл (так как это GGUF all-in-one)
|
| 94 |
+
chat_handler = CustomQwen2VLHandler(clip_model_path=model_path, verbose=True)
|
| 95 |
+
|
| 96 |
+
llm = Llama(
|
| 97 |
+
model_path=model_path,
|
| 98 |
+
n_ctx=8192, # Контекст (картинки большие, нужно место)
|
| 99 |
+
n_gpu_layers=0, # CPU
|
| 100 |
+
verbose=True,
|
| 101 |
+
chat_handler=chat_handler, # <-- ВАЖНО: Используем наш класс
|
| 102 |
+
n_batch=512,
|
| 103 |
+
logits_all=True
|
| 104 |
+
)
|
| 105 |
+
print("Модель успешно загружена с CustomQwen2VLHandler!")
|
| 106 |
+
except Exception as e:
|
| 107 |
+
print(f"Ошибка загрузки: {e}")
|
| 108 |
+
raise e
|
| 109 |
return llm
|
| 110 |
|
| 111 |
def process_image(image):
|
| 112 |
+
# Ресайз до 1024px макс, чтобы не перегружать CPU память и контекст
|
| 113 |
+
max_dim = 1024
|
| 114 |
+
if max(image.size) > max_dim:
|
| 115 |
+
image.thumbnail((max_dim, max_dim), Image.Resampling.LANCZOS)
|
|
|
|
|
|
|
| 116 |
|
| 117 |
buffered = io.BytesIO()
|
| 118 |
image = image.convert("RGB")
|
|
|
|
| 124 |
return "Пожалуйста, загрузите изображение.", ""
|
| 125 |
|
| 126 |
try:
|
| 127 |
+
progress(0.1, desc="Загрузка модели...")
|
| 128 |
model = load_model()
|
| 129 |
|
| 130 |
+
progress(0.2, desc="Обработка...")
|
| 131 |
+
base64_img = process_image(image)
|
| 132 |
+
img_url = f"data:image/jpeg;base64,{base64_img}"
|
| 133 |
|
| 134 |
system_prompt = "You are doing the image quality assessment task."
|
| 135 |
+
user_prompt = (
|
| 136 |
"What is your overall rating on the quality of this picture? "
|
| 137 |
"The rating should be a float between 1 and 5, rounded to two decimal places, "
|
| 138 |
"with 1 representing very poor quality and 5 representing excellent quality. "
|
|
|
|
| 144 |
{
|
| 145 |
"role": "user",
|
| 146 |
"content": [
|
| 147 |
+
{"type": "image_url", "image_url": {"url": img_url}},
|
| 148 |
+
{"type": "text", "text": user_prompt}
|
| 149 |
]
|
| 150 |
}
|
| 151 |
]
|
| 152 |
|
| 153 |
full_response = ""
|
| 154 |
+
print("Начинаю генерацию...")
|
| 155 |
|
| 156 |
+
# Запуск стриминга
|
| 157 |
stream = model.create_chat_completion(
|
| 158 |
messages=messages,
|
| 159 |
max_tokens=1024,
|
|
|
|
| 167 |
if "content" in delta and delta["content"]:
|
| 168 |
content = delta["content"]
|
| 169 |
full_response += content
|
| 170 |
+
yield full_response, "Думаю..."
|
| 171 |
|
| 172 |
+
# Поиск оценки
|
| 173 |
score_match = re.search(r'<answer>\s*([\d\.]+)\s*</answer>', full_response)
|
| 174 |
+
final_score = score_match.group(1) if score_match else "Оценка не найдена"
|
| 175 |
|
| 176 |
yield full_response, final_score
|
| 177 |
|
| 178 |
except Exception as e:
|
| 179 |
+
err_msg = f"Произошла ошибка: {str(e)}"
|
| 180 |
+
print(err_msg)
|
| 181 |
+
yield err_msg, "Error"
|
| 182 |
|
| 183 |
+
# Интерфейс
|
| 184 |
+
with gr.Blocks(title="VisualQuality-R1 (Custom Handler)") as demo:
|
| 185 |
gr.Markdown("# 👁️ VisualQuality-R1 (Qwen2-VL)")
|
| 186 |
+
gr.Markdown("Оценка качества изображений на CPU с кастомным обработчиком.")
|
| 187 |
|
| 188 |
with gr.Row():
|
| 189 |
with gr.Column():
|
|
|
|
| 192 |
|
| 193 |
with gr.Column():
|
| 194 |
output_score = gr.Label(label="Оценка")
|
| 195 |
+
output_text = gr.Textbox(label="CoT (Рассуждения)", lines=15)
|
| 196 |
|
| 197 |
run_btn.click(evaluate_image, inputs=[input_img], outputs=[output_text, output_score])
|
| 198 |
|